def setup_env():
num_concurrent_streams = Config.get("compiler", "cuda",
"max_concurrent_streams")
if 'ORT_USE_STREAMS' in os.environ:
ONNXRuntimeCUDA.use_streams = _env2bool(os.environ["ORT_USE_STREAMS"])
if ONNXRuntimeCUDA.use_streams:
log.info("Using streams with ORT (experimental)")
if num_concurrent_streams == 0:
log.info("Setting compiler.cuda.max_concurrent_streams to 8")
Config.set("compiler",
"cuda",
"max_concurrent_streams",
value=8)
elif num_concurrent_streams == -1:
ONNXRuntimeCUDA.use_streams = False
else:
if num_concurrent_streams != -1:
log.info("Setting compiler.cuda.max_concurrent_streams to -1")
Config.set("compiler", "cuda", "max_concurrent_streams", value=-1)
ONNXRuntimeCUDA.use_streams = False
ONNXRuntimeCUDA.max_concurrent_streams = Config.get(
"compiler", "cuda", "max_concurrent_streams")
def get_bool(self, *key_hierarchy):
from dace.config import _env2bool
res = self.get(*key_hierarchy)
if isinstance(res, bool):
return res
return _env2bool(str(res))